Online health monitoring framework of deep-sea UUVs via multi-sensor fusion and Koopman operator-based residual inference

Abstract

Deep-sea unmanned underwater vehicles (UUVs) depend on reliable propulsion systems during cruising, docking, and observation missions. However, online propulsion health monitoring remains challenging because high-frequency reference states are difficult to obtain in real time, while the nonlinear dynamics of UUVs make the relationship between thruster inputs and state responses difficult to characterize accurately. To address these challenges, this study proposes an online health monitoring framework that employs high-rate factor graph optimization (FGO) fused states as reference states and integrates UUV-specific Koopman modeling with persistent fault-evidence inference. To better capture the nonlinear dynamic characteristics of UUVs, task-specific state observables, nonlinear lifting functions, and delay embedding are designed to enhance the representation of nonlinear vehicle dynamics. On this basis, the prediction residuals are projected onto the thruster-input direction and further processed using a resettable cumulative likelihood ratio test to suppress transient disturbances and enhance persistent fault signatures. The proposed framework is validated through trajectory-tracking simulations and a 4000-m deep-sea experiment. The results demonstrate effective overall health monitoring, with fault warnings achieved within 25 s under multiple fault scenarios in the actual experiments. In addition, the method shows the potential to indicate the faulty thruster under specific operating conditions.